Improved co-registration of ex-vivo and in-vivo cardiovascular magnetic resonance images using heart-specific flexible 3D printed acrylic scaffold combined with non-rigid registration

Pre-clinical imaging experiments offer the opportunity for the validation of in-vivo imaging findings against high-resolution ex-vivo imaging examination. Ex-vivo cardiovascular magnetic resonance (CMR) combines the strengths of CMR tissue characterization without the limitation of tissue movement that occurs in-vivo and with greater flexibility for duration of acquisition while maintaining tissue 3D structural integrity. Ex-vivo CMR can achieve image resolution and quality surpassing that available during in-vivo CMR [1, 2] and has established the ability of in-vivo CMR to accurately characterize pathological processes. For example, ex-vivo CMR demonstrated the ability of late gadolinium enhanced (LGE) CMR imaging to differentiate viable from non-viable tissue in the chronic stage after myocardial infarction (MI) [3], that gadolinium accumulation differentiates myocardial scar from adjacent healthy tissue at a near cellular level [2] and has been used to accurately establish the area at risk (AAR) following experimental MI [4].

Despite these advantages, relating ex-vivo imaging findings to in-vivo observations is still associated with significant challenges. Within the thorax, the position of the heart is maintained by the great vessels and enclosure within the pericardial sac. The shapes of the chamber cavities and myocardium are affected by the internal cavity pressure and intrathoracic pressure as well as passive and active tension within the myocardial wall. When the heart is removed from the thorax it is no longer subject to these forces, which can result in a significant change in the shape of the heart, imposing a barrier to accurate registration back to in-vivo imaging, as shown in Fig. 1. In addition, tissue volumes change on transition to the ex-vivo condition, even prior to fixation, with implications for comparisons of tissue volumes between in-vivo and ex-vivo imaging [5]. Ideally, the comparison of structural features between in-vivo and ex-vivo imaging requires the accurate registration of in-vivo and ex-vivo data, which may be considered in two stages: The first step is maintaining the shape of the heart in the ex-vivo condition as close to its shape in the in-vivo condition as possible despite the changes in loading conditions and myocardial tissue state. The second step is the subsequent registration of the in-vivo and ex-vivo imaging data for direct comparison.

Various techniques have been proposed to overcome the tendency of the left ventricle (LV) to collapse when removed from the thorax, including inflating balloons with deuterated water within the LV cavity [6, 7] and injecting a setting resin into the LV cavity [8] or the coronary arteries [9, 10] prior to ex-vivo assessment. While effective at preventing LV collapse these techniques do not necessarily preserve the in-vivo shape of the LV cavity.

3D printing is a widely available technology and shows promise for the personalized assessment and management of a range of medical conditions as well as research applications. In clinical practice, 3D printed models offer incremental benefit over contemporary imaging when used in the planning of thoracic surgical procedures, which may reflect more effective transmission of imaging information to operators. 3D printed models have been used for optimization of prosthetic heart valve size selection [11] and to produce anatomically customized implants and prostheses in trauma surgery [12]. Furthermore, 3D printed MRI derived mounts for tissue analysis improve registration of CMR imaging to ex-vivo tissue samples when compared with conventional techniques [5, 13], facilitating accurate interpretation of histological specimens.

In this study, we developed and assessed a novel two-stage approach to optimize the registration of in-vivo and ex-vivo CMR in a porcine model of chronic MI. A flexible 3D printed scaffold based on a direct segmentation of in-vivo imaging in the shape of the LV cavity and aortic root was employed with the goal of preserving the intra-thoracic shape of the LV during ex-vivo CMR. Subsequently, a method for non-rigid registration of the ex-vivo imaging to the in-vivo imaging was applied. A comparison of the chamber volumes and the correspondence between areas of myocardial scar between the in-vivo and ex-vivo data was undertaken after each step. The results were compared with a cohort of experiments performed without the use of a 3D printed insert to maintain LV shape during ex-vivo imaging.

Segmentation of the in-vivo imaging for preparation of the scaffold took less than 10 minutes in each case. Refinement of the segmentation prior to printing was completed in each case in under 30 min and the printing was then completed within 180 min, followed by a further 30 min of post processing in each case. Ex-vivo imaging using a 3D scaffold was successfully acquired in all swine. An example of a 3D printed scaffold inserted into an excised heart is shown in Fig. 2. In 6 cases the scaffold was successfully passed across the MV. In a single case an incision at the basal infero-lateral wall was required because the MV annulus was too small to allow passage of the compressed scaffold. External examination demonstrated that the scaffold maintained the LV shape during ex-vivo CMR acquired over 4 h and there was low signal from the scaffold, which made identification of the endocardium unambiguous.

Myocardium was readily segmented from each of the ex-vivo scans and a region of increased signal intensity in the antero-septal wall was clearly visualized (example in Additional file 1: Figure S1). An example of corresponding positions within the myocardium on in-vivo imaging and ex-vivo imaging is shown in Fig. 3.

In this study, a two-stage co-registration process combining the use of a flexible, 3D printed scaffold and a non-rigid registration technique has been proposed and assessed to facilitate the accurate comparison between ex-vivo and in-vivo imaging. The main findings of this study are 1) the demonstration of non-uniform changes in tissue geometry that occur in the unloaded, ex-vivo condition and their quantification 2) the feasibility of using a flexible, 3D printed scaffold based on the in-vivo LV and aorta shape to prevent LV cavity collapse during ex-vivo scanning reducing LV shrinkage below 20% in each dimension and resulting in isotropic contraction of the LV cavity and myocardium, and 3) the demonstration of a subsequent non-rigid registration step to successfully correct for residual non-uniform changes in tissue geometry.

LV cavity shrinkage and wall thickening is clearly demonstrated in previous reports using ex-vivo CMR scanning [8, 23, 24], although comparisons of absolute volumes between the imaging techniques are infrequently reported. Tissue shape change in the ex-vivo state is the result of the removal of tethering structures, changes in loading conditions and muscle contraction due to the depletion of intracellular adenosine triphosphate (ATP), followed by an increase in the proportion of myocytes in the contracted state, prior to the onset of rigor mortis [25]. The proposed two stage approach enabled to fully correct for morphological changes occurring between in-vivo and ex-vivo imaging. Importantly, this framework is readily translatable to other in-vivo and ex-vivo imaging modalities [26, 27].

Optimizing the shape of the heart in the ex-vivo condition to best reflect the shape of the heart in the in-vivo condition was the first stage of this co-registration process. This step is key to ensure accurate non-rigid registration, which is more effective when initial geometric differences between the imaging data are minimised [28]. Data from the control experiments clearly demonstrated LV cavity collapse during ex-vivo imaging when no cavity support was used and the anisotropic contraction in this condition represents a significant obstacle which will reduce the accuracy of subsequent registration steps. Although in this experiment there was a delay of 1 week between in-vivo and ex-vivo data acquisition, the length of preparation of the scaffold would allow imaging to be acquired on the same day if required. An important consideration that informed the decision to use a flexible 3D printed scaffold was the ease with which the scaffold could be passed across the MV into the LV cavity. During preliminary experiments utilizing a rigid scaffold of the LV cavity [29], the maximum diameter of the LV occurred at a mid-ventricular level, and therefore in order to pass the scaffold across the MV annulus an incision was required in the MV annulus, as demonstrated in Additional file 1: Figure S2. In addition, due to the pseudo-symmetry of the LV cavity, the correct rotation of the scaffold around the LV long axis was uncertain. These experiences motivated the trial of a flexible 3D printed scaffold as assessed in this study. In all but one case using a flexible scaffold facilitated maintenance of LV cavity shape without incurring the disruption in tissue architecture that was required when using a rigid scaffold. The uncertainty of the rotation of the flexible scaffold was minimized by inclusion of the aortic root in the LV cavity segmentation, which was used a specific landmark to identify the correct rotation around the LV long axis, representing a significant benefit of using a flexible over a rigid 3D scaffold.

When specific tissue compartments such as scar or AAR are compared between the ex-vivo and in-vivo condition they are commonly reported as a percentage of LV myocardium and good agreement between in-vivo and the ex-vivo proportions are reported [6, 7]. Our data highlight that although the proportion of myocardium defined as scar prior to registration is consistent, there is less tissue thickening in the scar than the remote healthy myocardium, indicating that tissue shape change is non-uniform when the infarcted heart is removed from the thorax. In-vivo CMR imaging was acquired at end diastole, at which point active tension within the myocardium is minimal and the LV end-diastolic pressure (LVEDP) is approximately 10 mmHg [30]. The differential response to tissue unloading between healthy tissue and fibrous scar observed may be explained by differences in stiffness between scarred and healthy myocardium [31], and likely also reflects more pronounced myocardial contraction in regions of healthy tissue. These non-uniform tissue volume changes, if uncorrected, further compound the difficulty of comparison between in-vivo and ex-vivo data [32, 33]. The co-registration framework presented maintains LV structural integrity, minimizes shape change during ex-vivo data acquisition and fully compensates for residual non-uniform shape changes.

The volume differences of the LV cavity, LV myocardium and scar between the ex-vivo and in-vivo imaging despite the use of a scaffold reflect the morphological changes between the two conditions that represent a residual limitation to their accurate comparison, motivating the addition of the second non-rigid registration step in the co-registration process. The close correspondence of the compartment volumes following this step, as shown in Fig. 5 indicate accurate compensation of the volume changes between ex-vivo and in-vivo conditions. The DICE similarity coefficient calculated for the correspondence between regions of scar in the in-vivo and rigidly registered ex-vivo data demonstrate that the region of scar is well registered between the in-vivo and ex-vivo data sets following rigid registration alone (DICE coefficient = 0.75). Following non-rigid registration, the DICE coefficient for the scar rose to 0.77 which did not represent a statistically significant improvement (p = 0.550). This reflects the success of the strategy to maintain the shape of the LV using the 3D printed scaffold as well as the reduced morphological changes evident in the scar tissue versus the more compliant neighboring healthy myocardium. While successful in matching the location of the scar subjectively and with a reasonable DICE similarity coefficient, the DICE of the scar (calculated nodal-wise) was lower than that achieved with the myocardium and the LV cavity (calculated voxel-wise). The scale dependence of the DICE similarity coefficient [34, 35] with the smaller relative volume of scar, different signal thresholding strategies employed for scar segmentation in both the in-vivo and ex-vivo imaging, and the lower resolution of the in-vivo imaging which is limited to represent complex patterns of scar demonstrated on high resolution ex-vivo imaging may contribute to this. The use of more advanced rigid co-registration for improved pre-conditioning [36] or other non-rigid co-registration approaches [37] could also be investigated to further improve scar matching. The increase in absolute tissue volume in the ex-vivo condition remains unexplained. The most likely explanation is absorption of saline into the tissue following submersion during imaging. It is uncertain if the additional volume is due to expansion of myocytes due to disruption of the cellular membrane or additional fluid in the extracellular space and a histological assessment would be required to establish this.

Data from ex-vivo CMR has played a crucial role in the validation of in-vivo CMR imaging, facilitating its widespread adoption as a tool for assessing the accuracy of in-vivo structural and functional myocardial imaging [1, 2, 38]. There is an expanding role for ex-vivo CMR in the investigation of the local structural basis for observed physiological phenomena, where it has been used to establish thresholds for the interpretation of in-vivo electrogram voltage data [23], which contribute to the identification of appropriate ablation targets during ventricular tachycardia (VT) ablation [39]. Ex-vivo CMR is the principle modality used to generate high resolution biophysical models. These models have been used to predict successful ablation targets for the treatment of post-MI VT [40] and been proposed as a basis for interpolating clinical resolution imaging to generate higher resolution estimations of local scar architecture [41], with the goal of translating the insight from biophysical models to clinical ablation procedures [42]. The accuracy of these results depend on matching the global 3D structure of the ventricle in regions of healthy myocardium as well as scar. The proposed co-registration strategy was successful in matching the volume and shape of the LV cavity, myocardium and scar between in-vivo and ex-vivo data and may thus provide a novel avenue to improve the accuracy of these results with potential significant pre-clinical and clinical impact.

This study has several limitations. In the in-vivo condition the papillary muscles are within the LV blood pool, while in the ex-vivo condition they are pushed against the LV endocardial wall by the printed insert, because the segmentation method used for the printed insert does not account for portion of the blood pool between the papillary muscle and the endocardial LV surface in-vivo. This is likely to have introduced a discrepancy between the LV cavity shape in the in-vivo and ex-vivo conditions. We believe the impact of this would have been small due to the proximity of the papillary muscles to the endocardial surface in-vivo and note that there was not a clear indication of more pronounced shape changes in this region. A more complex in-vivo segmentation process for design of the printed insert could be considered in future to address this limitation of the study. The study involved comparing the LV morphology between the in-vivo and ex-vivo condition, however different imaging sequences were used for each. This permitted acquisition of higher resolution ex-vivo imaging data but introduced the confounding effect of differences in the imaging sequences to the assessment of registration. The impact of the different imaging sequences on the blood pool and myocardial segmentation should be minimal as these are unambiguously defined in both data sets. Furthermore, prior to registration the ex-vivo imaging was resampled to the resolution of the in-vivo imaging data to minimize the impact of differences in image resolution on the comparison between the data sets. The impact of the differences in imaging sequence on scar segmentation is more challenging to establish. In addition, no consensus exists for the best strategy to threshold ex-vivo CMR images for the identification of scar. A simple and objective strategy for scar thresholding in the ex-vivo condition was selected to maintain consistency but may have contributed to small differences in scar assessment between the in-vivo to ex-vivo conditions. The correlation between scar volume between the in-vivo and ex-vivo data indicates that this resulted in relatively minor differences between tissue identified as scar between the two data sets, but the overall impact of differences in the imaging sequences acquired in the ex-vivo and in-vivo conditions remains a potentially significant confounding factor. Despite the printed scaffold, there was significant shrinkage of the LV cavity and tissue thickening in the healthy myocardium prior to the non-rigid registration step. We hypothesize that pressurizing the flexible 3D printed scaffold to reach LVEDP and the application of an excitation-contraction uncoupler [43] may further reduce the LV cavity reduction during ex-vivo imaging and represent a potentially valuable avenue for future work. There was no direct comparison between ex-vivo imaging acquired with and without the 3D printed scaffold, due to the time limitation during which ex-vivo imaging can be acquired, but the comparison with separately acquired ex-vivo imaging without the use of a scaffold indicates the improvement in the relationship between myocardial and LV cavity volumes resulting from the use of the scaffold. Finally, the process of segmentation and image registration uses three separate computational libraries, each chosen for the advantages offered by specific features within the library. While successful, this has resulted in a complex process of registration using multiple tools. This process could be simplified in future studies by development of a single interface with access to multiple libraries.

The pattern of the morphological changes seen between the in-vivo and the ex-vivo LV differ between scar and healthy myocardium. A two-stage approach to optimize the co-registration of ex-vivo to in-vivo imaging to facilitate accurate comparison is reported. A 3D printed flexible scaffold based on the in-vivo shape of the LV cavity is an effective strategy to reduce morphological changes in the ex-vivo LV. This was combined with a non-rigid registration approach to reduce residual morphological changes, which further improved the co-registration and local comparison between in-vivo and ex-vivo imaging data.